The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents a 'self-audit' mechanism in audit-loops.md and audit-log.md that directs the agent to evaluate its own performance and 'auto-apply' modifications to its core configuration files like SKILL.md. This instruction to self-modify poses a risk as it allows for the dynamic alteration of the agent's underlying rules and could be used to bypass security constraints or persist malicious logic.
[INDIRECT_PROMPT_INJECTION]: The audit templates provided in audit-loops.md (e.g., for the 'Naive Viewer' lens) ingest external video scripts and narration text. The templates interpolate this untrusted data into new prompts without using boundary markers or instructions to ignore embedded directives, leaving the agent's internal reasoning process vulnerable to manipulation by the content it is processing.
[EXTERNAL_DOWNLOADS]: The skill recommends the use of standard development tools and libraries, including the remotion framework and @ffmpeg-installer/ffmpeg for frame extraction. It also makes use of well-known AI generation services from providers such as OpenAI, Google, and xAI. These references are consistent with the skill's stated purpose and use reputable sources.

video-content-strategy